| Time-series classification is a very valuable task in data mining,which is widely used in different fields such as medical treatment,finance,power and signal processing.Therefore,the study and exploration for time-series classification is always a concerned task by researchers.With the explosive growth of time-series data scale,the classification method based on statistical theory has already reached the bottleneck of accuracy;The calculation cost of classification methods based on traditional machine learning is too high,while classification methods based on depth learning are difficult to have good generalization performance.Most of the methods are aimed at the task of single-variable time-series classification,while the problems of multi-variable timeseries classification are more often encountered in reality.The time-series classification faces new challenges.For single variable time-series classification task,we propose a time-series classification method based on wavelet transform and convolution neural network,called Wave CNN.The time-series datum are converted into time-domain and frequency-domain representations by wavelet transform,and then features are extracted from these representations through convolution neural networks and customized pooling operations,which can greatly reduce the training time and enhance the generalization performance of the model.The experimental data consists of 109 data sets of UCR Repository,and in this work,Several popular models are selected as comparison model to conduct experiments about time-series classification.The experimental results show that the classification accuracy of Wave CNN proposed in this paper has reached the equivalent level as the current most advanced models,at the same time,the training time of Wave CNN is at least one order of magnitude faster than other models with the equivalent accuracy.For multivariate time-series classification tasks which are less studied and explored,we propose an ensemble multivariate time-series classification method that contains several different modules which can reduce the dimension of time-series datum.Each module will independently train a Wave CNN classifier.The final prediction result is formed by the prediction results of each module using CAWPE.The experimental data consists of 26 data sets of UEA Repository,experimental results show that Ensemble WC is ahead of all other similar methods in several evaluation indicators.By expanding of training time,ensemble has improved the classification accuracy and enhanced the generalization performance of WaveCNN. |
